Sometimes it’s easy to know which messages will spread through Twitter like wildfire. Just ask Rep. Anthony Weiner (D-New York), who faces pressure to resign after unwittingly sending an intimate photo of himself to thousands of followers.
Researchers at MIT’s Laboratory for Information and Decision Systems are testing a search engine that identifies which posts on a given topic are likely to spread by studying the network of connections between users.
The system, called Trumor, identifies people who are well-positioned to spread information, and uses this to weight the value of different posts on a given topic. Information usually spreads between users as they “retweet” posts. To find influential Twitter users whose posts will likely be retweeted, the researchers examined the network of tweets and retweets around topics such as tennis, soccer, and the BET Awards. Early results suggest the technique could provide an effective way to find posts that will spread broadly through the network.
Automatically identifying influential Twitter users could be useful to advertisers, who could use it to spread information about products more effectively.
Determining influence on Twitter isn’t as simple as seeing how many followers a user has. What matters most is that followers pay attention to posts and discuss them, and that this discussion spreads beyond the user who started it. The researchers have been exploring better ways to measure a person’s influence, and Trumor grew out of this work.
The team began by studying networks of retweets on Twitter. They grouped retweets by topic and looked at how they spread through the network. The researchers considered users to be connected in the network if one retweeted a message from the other—simply following each other wasn’t enough.
Once they had those networks, a clear pattern emerged, says Tauhid Zaman, a PhD candidate at MIT’s Laboratory for Information and Decision Systems who was involved with the work. For each topic, they found “superstars”—highly connected individuals whose posts spread widely. The influence of these people was far greater than that of others within their network.
In many cases, it would be useful to be able to identify these users before an event takes place. For example, an advertiser might want to talk with someone before the BET Awards to get out information about a product during the event.
The researchers tested several methods of doing this, such as looking at the number of connections a user had, or how close they were to others on the network. They found that they could identify them using a method called “rumor centrality,” which measures how well-placed a person is to spread information. The technique measures how many paths a user has for spreading information.
Zaman says that rumor centrality is particularly valuable because it takes the entire network into account, not just the connections in the user’s immediate vicinity. For example, a person might have a lot of followers, but those followers might not be well-connected themselves. A person with fewer, better-connected followers has more paths for spreading information, and therefore a higher “rumor centrality” score.
Once they found a method of identifying superstars, the researchers built an experimental search engine around the system. Trumor finds people with high rumor centrality scores for a given topic and weights their posts, yielding pieces of information that are most likely to spread. Users can select a topic they want to search and be directed to pieces of information that could prove popular. The system does identify popular accounts, such as that of Lady Gaga, but, Zaman adds, it also pulls up relative unknowns. He says Trumor is still in its early stages, but adds that tests suggest it does well at identifying timely, pertinent information.
Other researchers are also looking at ways to measure influence on social networks. Abhik Das at the University of Texas at Austin has conducted studies on influence on cellular phone networks, and found that the structure of a network as a whole is a key factor. But his work also suggests that a person’s influence waxes and wanes over time, and that a good system must take this into account. “A person can’t go on spreading influence indefinitely,” Das says.
Zaman agrees, and says the plan is for future versions of Trumor to calculate rumor centrality for a window of time, such as the past week or month, allowing changes in the network to affect how information is weighted.